Evidence standards for multi-ancestry polygenic prediction

Polygenic scores (PGS) are prospective tools for health screening, prevention and trials, but most are trained in European genome-wide association studies and lose accuracy and calibration in non-European populations and those of mixed ancestral heritage. Multi-ancestry methods are proliferating, yet benchmarking standards lag. This review focuses on benchmark design for cross-population prediction. We show how ancestry assignment, linkage disequilibrium references, variant sets, tuning and trait architecture shape apparent performance. We propose a scorecard spanning discrimination, calibration, equity gaps and compute cost, plus stress tests in diverse cohorts and realistic simulations. Finally. we outline a future vision of living, auditable benchmarking frameworks.